Answers are grounded in this paper and cite their sources.
The selection of papers related to a given topic involves forming an optimal order of the user working with primary sources from more general to more specific. In the ideal case, we obtain an estimation for the mutual semantic dependence of texts relative to the most rational (i.e., standard) variants of description for knowledge fragments represented by them. Substantially, here, we have the problem of the splitting of texts into classes by means of their clustering with center of mass revelation for each cluster. In the current study, the mentioned problem is solved by application of neural network language models of the Transformer architecture. The main idea of the proposed solution consists in the splitting of analyzed texts (i.e., abstracts of scientific articles together with their titles) into classes according to semantic similarity by means of clustering of multi-dimensional vectors called embeddings corresponding them. In this study, variants of clustering by applying k-means (substantially, k-means++), HDBSCAN, and spectral clustering methods are implemented. Within a cluster, an annotation with an embedding closest to the center of the cluster (corresponding to the center) is considered to be closest to the reference (i.e., standard) variant of conveying meaning among the annotations of the cluster and, accordingly, the starting point in the navigation trajectory through the articles related to the cluster. In the case of two clusters, the navigation trajectory through the collection starts from the found center of one cluster and then proceeds to the found center of the other cluster in descending order of the semantic coherence of the corresponding annotations. In the case of three and more clusters, to construct the final trajectory of user navigation across the collection, those abstracts, which are correspond to centers of clusters, are sorted in descending order of the product of estimation of the semantic coherence and estimation of the “strength” of the semantic relationship with mass centers of other clusters. Herewith, the estimation of the strength of the semantic relationship of abstracts is defined by the measure of closeness of corresponding vectors based on cosine similarity. The semantic coherence of the abstract here indicates how much its sentences are related to each other in meaning. Here, the estimation of abstract’s semantic coherence is substantively close to the estimation of the “strength” of the publication’s semantic relation with other papers in the collection and can be calculated based on the cosine similarity of the corresponding embedding vectors for sentences of the analyzed abstract.
No takes yet. Share an insight, caveat, or question.
Mikhaylov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: